Oliver Galgut, Elselijn Kingma (University of Southampton)
Introduction: Random allocation (randomisation) is widely considered the best allocation technique for double-masked, controlled, interventional medical trials. One problem for randomisation is that — contrary to what is often claimed — it cannot perfectly control for all potential confounders. Due to chance alone, important variables can be differentially distributed at baseline. Dynamic Allocation (DA) is a potential alternative to randomisation, developed as a response to this problem of baseline imbalance. DA techniques actively allocate patients to ensure the treatment and control groups are as similar as possible. This ensures that treatment effects are unlikely to be confounded. Unfortunately, DA techniques are not widely used because of fears that they may be susceptible to bias. In this poster, we examine the ability of DA to prevent bias and achieve balance.
Methods: We shall assess DA by comparing it to randomisation using two key justifications of randomisation:
1. Randomisation prevents bias
2. Randomisation is the only method that guarantees that all confounding factors (known and unknown) can be balanced between treatment and control groups.
These two justifications were chosen because they are foundational to randomisation’s justification. By examining DA’s performance on these, we can consider whether DA is a serious potential competitor to randomisation.
Results: DA is better at balancing known confounders than randomisation. It is no worse than randomisation at balancing unknown confounders. DA is demonstrated to prevent biases, independent of balancing ability. Additionally, we find that the ideal DA technique: 1) inputs continuous covariates, 2) uses a variable allocation probability, and 3) is opaque.
Conclusions: DA can be at least as good as randomisation at preventing bias and achieving balance. Therefore, it is a competitor to randomisation. This is particularly true if the technique inputs continuous covariates, uses variable allocation probabilities, and is opaque. Such techniques should be considered on an equal footing to randomisation when designing interventional trials.
